Hark Handoff: Low-Cost Computer-Use Agent Explained

Hark Handoff is Hark's first product: a computer-use agent that navigates websites and takes actions for the user. It can reportedly order food, book flights, and message candidates on LinkedIn. The practical pitch for builders is straightforward: instead of integrating with every service API, an agent can operate the browser directly. That matters because many websites still do not offer a usable public API.

Handoff treats the browser as the integration layer

For each request, Handoff reportedly creates a dedicated virtual computer with its own browser, file system, and terminal. Users can connect existing accounts, allowing the agent to use saved addresses, payment methods, and account history. This architecture gives the agent more room to complete end-to-end tasks than a chatbot that only returns instructions.

For an AI product team, the trade-off is control. A virtual machine can isolate a task and preserve the browser state it needs, but it also creates operational questions around credentials, files, permissions, audit logs, and human approval before irreversible actions. Hark has not yet explained who can access those computers or the files stored on them.

The Hark computer-use agent makes an aggressive performance claim

Hark says Handoff scored 97.7 on Online-Mind2Web, ahead of GPT-5.4 at 92.8, Claude Opus 4.8 at 84.1, and Gemini 2.5 Pro at 69. The result is significant if the evaluation setup is comparable, but it does not establish that Handoff is the best web agent available.

The comparisons omit newer models including GPT-5.6 and Opus 5. Hark also measured some results in its own harness and used an internal language model judge for certain evaluations. On WebTailBench v2, GPT-5.5 reportedly scored 72.3 compared with Handoff's 68.6. Builders should therefore treat 97.7 as a vendor-reported benchmark result, not a complete product verdict.

Handoff pricing could matter more than the leaderboard

Hark states that Handoff pricing is $0.18 per million input tokens and $2.37 per million output tokens, with roughly 0.8 seconds of per-turn model latency. Those figures would be substantially below the cited frontier model prices and could reduce costs for long agent runs, where repeated screenshots, actions, and reasoning accumulate quickly.

That advantage still needs to be tested against the full workflow cost. Token charges do not necessarily include virtual computer infrastructure, browser sessions, storage, retries, supervision, or failed transactions. For production teams, successful task completion and review burden matter more than model price alone.

Availability and security remain deployment blockers

Hark has described Handoff as a technical preview and indicated availability in late summer 2026, while saying that security and privacy are a primary focus. The company has not supplied independent audit, retention, access-control, or enterprise data-handling details in the provided evidence.

Hark raised a $700 million Series A at a reported $6 billion valuation, led by Parkway Venture Capital with participation from Nvidia, AMD, Intel Capital, Qualcomm Ventures, Salesforce Ventures, and ARK Invest. Brett Adcock remains CEO of both Hark and Figure AI, a background that has attracted scrutiny around public claims about Figure's BMW partnership, but that context does not by itself validate or invalidate Handoff's technical results.

The sensible decision rule for builders is to watch Handoff, but begin with reversible tasks. Its low token price and browser-native design could be useful for recruiting, research, and back-office automation. Do not give it payment credentials or sensitive enterprise files until Hark publishes clear isolation, logging, retention, and human-approval controls.